The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill uses Python's 'pickle' module for local caching and checkpointing of protein embeddings and generation results. The 'EmbeddingCache' class in 'references/esm-c-api.md', and the 'CheckpointedBatchProcessor' and 'ForgeCache' classes in 'references/forge-api.md' all utilize 'pickle.load()' to deserialize data from local files ('embeddings_cache.pkl', 'checkpoint.pkl', and files within 'forge_cache/'). Because 'pickle' can be used to execute arbitrary code during deserialization, these files represent an attack vector if an adversary can modify them.
[EXTERNAL_DOWNLOADS]: The skill fetches model weights and connects to the EvolutionaryScale Forge API for inference. It references domains such as 'forge.evolutionaryscale.ai' and 'github.com/evolutionaryscale/esm'. These are official vendor resources for this skill.
[DATA_EXPOSURE]: The documentation provides guidance on using API tokens for the Forge platform. While examples show placeholders like '', the 'references/forge-api.md' file correctly recommends managing these secrets via environment variables or a secrets manager, which aligns with security best practices.
[PROMPT_INJECTION]: The skill processes protein data from external files (e.g., '.pdb' files via 'ESMProtein.from_pdb()'). While this presents an ingestion surface for potentially malicious data, the specialized nature of the protein models (ESM3, ESM C) significantly limits the risk of traditional instruction-based injection compared to standard LLMs. Evidence: 'ESMProtein.from_pdb()' in 'references/esm3-api.md', 'SKILL.md', and 'references/workflows.md'. Capability: File writing and Network operations via Forge API.

esm